Tool call testing

[alay2shah]

This PR adds tool call format guardrails to the training pipeline.

• Detects LFM2, LFM2.5, structured (OpenAI), and foreign formats from a sample of the data
• Auto-strips <|tool_response_start|> from role=tool messages to prevent double wrapping (the LFM2 template adds these itself)
• Auto-strips <|tool_list_start|> when training LFM2.5 models
• Auto-converts OpenAI-style tool_calls fields to pre-baked bracket notation
• Hard errors on foreign formats (<tool_call> XML etc.) with a message telling you exactly what to do
• Warns on format mismatches it can't auto-fix (e.g. missing <|tool_list_start|> for LFM2)

Pipeline integration is light. The normalization runs as a Ray .map() step between column normalization and filtering — same pattern as existing transforms. Existing files got ~10 lines each. model_name flows from the YAML config through DatasetLoader so the pipeline knows which model family to validate against.

Also re-enables preprocess_fn — users can now specify preprocess_fn: "my_module.my_func" in their YAML config for custom dataset transforms.

[Rouzbehat78]

I did add few edge cases in the pytests where the tool_calls_to_pythonic converter was outputting invalid strings that arent parsable when a special character existed in the string like (", , \n, etc)
I also pushed the small fix, so we use json dumps for those cases that will make sure it's parsable stings.

[Rouzbehat78]

Are we using this anywhere?

[alay2shah]

Agents sometime use it but it's relatively harmless alone

[alay2shah]

Lgtm, just approve